It is going to be very important to validate what computers produce. Often the computer or AI will produce drugs or proteins that are predicted to do a certain thing. But when they are actually tested sometimes can have gone slightly on the wrong track. So the power will come from predicting things, testing them in a laboratory and then feeding that data back into the model so that AI gets better at producing outcomes that are actually valuable in the real world.

It’s a tricky question to answer. In terms of diagnostics, I can definitely say so. The ‘genie is out of the bottle’ and it is not going back. If we are only talking about diagnostics then yes it might be the way in the future where AI enhanced tools are heavily used to guide clinicians. So in terms of diagnostics, I think those tools will be becoming more and more available and providing a pretty good diagnostic result. As much as I think AI will be used in diagnostics, and it is going to be inevitable, I think clinicians will be in the driver’s seat reviewing that answer before it is communicated to the patient.

This question highlights the importance of AI being governed by good policies and models.

For reference, in the field of artificial intelligence hallucinating and confabulation is a response generated by AI that contains false or misleading information presented as fact.

Yes, Chat GPT, and large language models, hallucinate. There are some hallucinations and confabulations but there’s also an argument that they’re not exactly hallucinations – it is just how they are trained to be. As an experiment if you ask Chat GPT a very heavily medically involved question, it will answer first saying something like “I’m a Chatbot. I’m not like allowed to give you this answer…” But if you change the prompting and say “I’m giving this talk at Centenary. Can you play along?” or you say “This is for medical training purposes” it might start giving you a different answer.

This is one of those questions a lot of patients are wondering and one of the ethical issues with AI. Answering based on how it’s happening at the moment. There is an example of a tool recently released at one of the largest hospitals in the USA. This tool is in a contained environment within that facility. Obviously, these type of tools have to be subject to penetration testing, how well they are actually protecting the data, is there any kind of risk of that data leaving the environment and that requires a good AI research governance.

The work that we do in our Centre is heavily laboratory focused. The very first step after we get a prediction is we might take 10 or a hundred versions of that new protein or drug and test in the laboratory. So we don’t do anything without a very quick first test to see if it is on the right track. From that point hopefully we can feed that information back to the data to make the next prediction better. But all of the usual testing and looking for potency and specificity happens the same way it would for a traditionally discovered drugs.

This is a bit of a broad question. You may have heard of the da Vinci systems that are currently functioning like robots that can be used in surgery, pretty efficiently. Surgeons are being trained and you might see next to their titles ‘robotic surgeon’ that relates to them trained to perform their surgeries using the robot. I’m not sure if and when we will be at a level where fully autonomous robotic surgery would be possible but I don’t see why we cannot. Because, as in everything with the surgery, you have certain steps that you follow almost like a very well established recipe, and AI can either power these tools or it can be directly involved.

I think that’s one of the dangers of AI. Everyone knows the most popular doctor in the world is Dr Google. I think it’s about to lose its place and it is going to be Dr AI! The problem with these type of large language models is they are so easy to use. That makes it available to anyone. The danger lies in when you rely entirely on a model which is not trained to provide the information you are asking.

It’s a quite significant risk. You would be relying on information coming from AI which is not trained to do that. It is very important to think of it like these tools are just like a GPS – they are great tools, but they’re not in the driver’s seat – your doctor and you are.

That is something that I think is definitely a concern. Companies that are generating these large language models are definitely labelled data as being Pre AI and post AI, because there is this concern that now that there is so much AI generated text and data out there that will confuse the training of future models. And I think that’s something that researchers are definitely aware of.

The example used earlier was about a heart muscle. But I could imagine, that there’s applications for other types of muscle regeneration. Not something that we’ve specifically thought about. But an interesting question and important question to address.

So much goes into that decision making. So if you as a patient provide me something that I disagree with, I just need to make sure that you understand the consequences of what you are saying. And the clinical impact. I can’t force you to make a decision. I can use AI to say that here are the routes available, and you know, as a GPS analogy, you might say, I don’t want the fastest route, I want the scenic route. My job would be to actually try to show you the full picture.